Evidence correction, 14 October 2026: earlier versions of this post described API prices as below-cost subsidies, which reported compute margins do not support, misstated OpenAI's 2026 revenue and spending, said Google cut Gemini prices when it raised them, misdated the xAI government deal, and credited unnamed analysts with 3x to 10x price-increase estimates that no source supports. They also misreported the coding-agent billing changes, an API reseller's cost comparison, Sora's revenue, the routing-layer funding total and Portkey price, two model prices, a FinOps survey finding, and one routing calculation. Those claims are removed or corrected below.
The number that should change how you think about your AI bill.
OpenAI projects $14 billion in losses for 2026. Not revenue. Not spending. Losses. The figure comes from internal projections that The Information first reported in October 2024, and it is still repeated in 2026 coverage of the company's finances.
Source: Yahoo Finance, "OpenAI's own forecast predicts $14 billion loss in 2026," January 2026
Those losses are driven by training compute and by the cost of serving hundreds of millions of free users. Paid usage looks different. The Information reported that OpenAI's compute margin, the share of revenue left after the cost of running models for paying users, reached about 70% in October 2025. So your API bill is not a documented subsidy. The risk is different: a company losing money at this scale may change its prices, and some analysts argue that prices will stabilize or rise.
Source: Fortune, "OpenAI sees better margins on business sales, report says," December 2025
This is not a short-term dip. The same projections show $44 billion in cumulative losses from 2023 through 2028 before the company expects to turn a profit in 2029. HSBC analysts concluded in November 2025 that OpenAI likely will not make money by 2030 and still faces a $207 billion funding shortfall to power its growth plans.
Just over 5% of ChatGPT's roughly 900 million weekly users pay. Sam Altman said in January 2025 that OpenAI was losing money on $200-per-month ChatGPT Pro subscriptions. When OpenAI announced in March 2026 that it was shutting down Sora, Forbes had estimated the app could cost as much as $15 million a day to run, while Appfigures counted about $1.4 million in consumer spending on it through January 2026.
Source: Fortune, on Sam Altman's comment that OpenAI was losing money on ChatGPT Pro, January 2025
Source: TechCrunch, "OpenAI's Sora app is struggling after its stellar launch," January 2026
Source: Engadget, "OpenAI is shutting down its Sora video generation app," March 2026
The prices you pay today are set by companies that are still losing money overall. That makes them a business decision that can change, not a fixed cost.
Each provider is running a different play.
The pricing moves across providers point in different directions.
Google launched Gemini 3.5 Flash at I/O on May 19, 2026, at $1.50 per million input tokens and $9.00 per million output. That is three times the price of Gemini 3 Flash Preview and 25% below Gemini 3.1 Pro ($2/$12). Simon Willison read the launch as the major labs starting to probe how much their API customers will pay.
Source: Simon Willison's Weblog, on the Gemini 3.5 Flash launch, May 2026
Under a GSA OneGov agreement announced in September 2025, US federal agencies can use Grok 4 for $0.42 per agency for 18 months. That is a rounding error, not a price.
Source: GSA news release on its OneGov agreement with xAI, September 2025
Anthropic is the one partial exception to the losses. The company confidentially submitted a draft S-1 to the SEC on June 1, 2026, days after a funding round at a $965 billion post-money valuation. The Wall Street Journal reported that Anthropic projected its first profitable quarter in Q2 2026, with $559 million in operating profit on $10.9 billion in revenue, and expected compute costs to fall from 71 to 56 cents per dollar of revenue. Its run-rate revenue crossed $47 billion in May 2026.
Source: Anthropic, "Anthropic confidentially submits draft S-1 to the SEC," June 2026
Source: Anthropic, Series H announcement, May 2026
Source: Bloomberg Law, "Anthropic Raises at $965 Billion Valuation, Eclipsing OpenAI," May 2026
The point is the same regardless of which provider you use: the current price level is a business decision, and providers are already changing it in both directions. Google raised its Flash price while DeepSeek cut V4 Pro by 75%. Neither direction is guaranteed to hold.
What happens if prices rise.
Some in the industry expect prices to stop falling, or to rise.
Arcade.dev's December 2025 analysis of inference economics argued that inference pricing will stabilize or increase within 18 to 36 months, because current prices reflect a race for market share rather than sustainable unit economics.
Source: Arcade.dev, analysis of AI inference economics, December 2025
An increase does not have to be dramatic to be painful. A 2x increase in frontier API pricing would double the bill of every team running always-Opus or always-GPT-5.5 workflows. For enterprise teams spending $50,000 to $100,000 per month on inference, that is a $600,000 to $1.2 million annual increase.
Meanwhile, coding agents are moving to token-based billing. OpenAI moved Codex from per-message credits to token-based credits on April 2, 2026. GitHub moved every Copilot plan to usage-based billing on June 1. Anthropic planned to move Agent SDK and headless Claude Code usage to a separate monthly credit on June 15, then paused that change the day it was due and called the pause temporary.
Source: OpenAI Help Center, "Codex rate card"
Source: GitHub Changelog, "GitHub Copilot is moving to usage-based billing," April 2026
Source: DevOps.com, on Anthropic pausing the Agent SDK billing change, June 2026
The pattern is consistent: as agentic workloads consume far more tokens than chat, providers are making the per-token cost visible.
The correction math for a typical team.
Consider a team spending $30,000 per month on LLM inference today, running everything through a frontier model at $5/$25 per million tokens (input/output). Assume, for illustration only, that routing cuts this team's bill by half.
| Scenario | Monthly bill | Annual cost |
|---|---|---|
| Current rate | $30,000 | $360,000 |
| 2x price increase | $60,000 | $720,000 |
| 3x price increase | $90,000 | $1,080,000 |
| Current rate + routing (hypothetical 50% savings) | $15,000 | $180,000 |
| 3x increase + routing (hypothetical 50% savings) | $45,000 | $540,000 |
The last row is the critical one. With routing, this team pays $15,000 per month at today's prices. After a 3x increase it pays $45,000 ($90,000 x 0.5), still above today's unrouted bill but half of the $90,000 an unrouted team would pay.
Under these assumptions, routing does not remove the risk of a price increase. It halves the exposure: the gap between today's bill and the 3x case shrinks from $60,000 to $30,000 a month. The 50% is an illustration, not a forecast. Nadir publishes no universal savings rate, so measure yours on your own traffic.
The optimization surface is wider than you think.
The savings from routing come from a simple observation: most API calls do not need a frontier model. Datadog's State of AI Engineering 2026 report measured production telemetry across thousands of companies and found that 69% of all input tokens are system prompts, tool schemas, and policy definitions that repeat on every call.
Source: Datadog, "State of AI Engineering 2026"
AI.cc, a Singapore-based API aggregator, said in a May 2026 press release covering 2.4 billion API calls on its platform that customers routing across model tiers paid a median $2.31 per million tokens in 2026, against $18.40 for frontier-only workloads twelve months earlier. That comparison mixes routing with a year of price cuts, so it does not isolate what routing alone saves.
The tier spread between models is wide. Claude Haiku 4.5 costs $1/$5 per million tokens. Claude Opus 4.8 costs $5/$25. GPT-5.4 costs $2.50/$15. DeepSeek V4 Pro costs $0.435/$0.87. On output tokens, Opus 4.8 costs about 29 times as much as DeepSeek V4 Pro ($25 / $0.87) and five times as much as Haiku 4.5 ($25 / $5). That gap is structural. Smaller models cost less to run, and that does not change if list prices rise.
When prices rise, the absolute savings from routing grow proportionally. A hypothetical 50% reduction on a $25 output rate saves $12.50 per million output tokens. The same reduction on a $75 output rate saves $37.50. Routing gets more valuable as prices rise, not less.
Model-agnostic architecture is the hedge.
The safest enterprise strategy is to build model-agnostic workflows today so that switching providers or moving to local inference is an operational decision rather than a re-engineering project.
Source: Arcade.dev, analysis of AI inference economics, December 2025
This is what a routing layer provides. Instead of hardcoding a single provider and model into every call site, you route through a classification layer that matches each request to the cheapest model that can handle it. When prices change, you adjust the routing thresholds. When new models launch, you add them to the pool. When a provider has an outage, you fail over to the next tier.
The FinOps Foundation surveyed 1,192 respondents representing more than $83 billion in annual cloud spend and found that 98% of FinOps teams now manage AI costs, up from 31% two years ago. Respondents named visibility into AI costs as a top challenge, since pricing models vary widely across providers. A routing layer with per-request analytics closes part of that gap.
Source: FinOps Foundation, "State of FinOps 2026"
IDC predicts that by 2028, 70% of top AI-driven enterprises will use dynamic model routing. The market moved the same way: in May 2026, OpenRouter raised $113 million at a $1.3 billion valuation, and Palo Alto Networks closed its $117 million purchase of the AI gateway Portkey.
Source: IDC, "The future of AI is model routing," November 2025
Source: TechCrunch, "OpenRouter more than doubles valuation to $1.3B in a year," May 2026
What to do before prices move.
Nobody knows whether or when list prices will rise. It could be 6 months, 12 months, or 24 months, or it may not happen at all. OpenAI cannot lose $14 billion a year indefinitely, and pricing is one of the levers it has.
Three steps that take less than a day and reduce your exposure:
1. Audit your model usage. Pull your API logs and count how many requests go to frontier models versus cheaper alternatives. Then count how many of those calls are simple classification, formatting, file reads, or boilerplate that a $1/M token model can handle as well as a $5/M token model.
2. Add a routing layer. Route simple requests to cheap models, mid-complexity to Sonnet-class models, and only complex reasoning tasks to frontier models. The blended cost drops immediately, and you gain the architectural flexibility to adjust when prices change.
3. Set up per-request cost tracking. If you cannot see cost per request, you cannot optimize. Tag each request with the model used, tokens consumed, and estimated cost. When a price change lands, you will know the impact in minutes instead of waiting for the monthly invoice.
The teams that treat current API prices as permanent are building on an assumption. The teams that treat routing as infrastructure are building a hedge. When prices change, the first group scrambles. The second group adjusts a config.
Sources: [The Information, "OpenAI Projections Imply Losses Tripling to $14 Billion in 2026"](https://www.theinformation.com/articles/openai-projections-imply-losses-tripling-to-14-billion-in-2026) (October 2024). [Yahoo Finance, "OpenAI's own forecast predicts $14 billion loss in 2026"](https://finance.yahoo.com/news/openais-own-forecast-predicts-14-150445813.html) (January 2026). [Fortune, "OpenAI sees better margins on business sales, report says"](https://fortune.com/2025/12/21/openai-compute-margins-revenue-chatgpt-earnings-sam-altman/) (December 2025). [Fortune, "OpenAI won't make money by 2030 and still needs to come up with another $207 billion to power its growth plans, HSBC estimates"](https://fortune.com/2025/11/26/is-openai-profitable-forecast-data-center-200-billion-shortfall-hsbc) (November 2025). [9to5Mac, "ChatGPT user base surges 350% in 18 months"](https://9to5mac.com/2026/02/27/chatgpt-approaching-1-billion-weekly-active-users/) (February 2026). [Fortune, Sam Altman on ChatGPT Pro losses](https://fortune.com/2025/01/07/sam-altman-openai-chatgpt-pro-subscription-losing-money-tech) (January 2025). [Forbes Australia, "OpenAI could be blowing as much as $15 million per day on silly Sora videos"](https://www.forbes.com.au/news/innovation/openai-could-be-blowing-as-much-as-15-million-per-day-on-silly-sora-videos/) (November 2025). [TechCrunch, "OpenAI's Sora app is struggling after its stellar launch"](https://techcrunch.com/2026/01/29/openais-sora-app-is-struggling-after-its-stellar-launch/) (January 2026). [Engadget, "OpenAI is shutting down its Sora video generation app"](https://engadget.com/ai/openai-is-shutting-down-its-sora-video-generation-app-211023358.html) (March 2026). [Simon Willison's Weblog, on Gemini 3.5 Flash](https://simonwillison.net/2026/May/19/gemini-35-flash/) (May 2026). [GSA news release on its OneGov agreement with xAI](https://www.gsa.gov/about-gsa/newsroom/news-releases/gsa-xai-partner-to-accelerate-federal-ai-adoption-09252025) (September 2025). [Anthropic, "Anthropic confidentially submits draft S-1 to the SEC"](https://www.anthropic.com/news/confidential-draft-s1-sec) (June 2026). [Anthropic, Series H announcement](https://www.anthropic.com/news/series-h) (May 2026). [Bloomberg Law, "Anthropic Raises at $965 Billion Valuation, Eclipsing OpenAI"](https://news.bloomberglaw.com/private-equity/anthropic-raises-at-965-billion-valuation-eclipsing-openai-1) (May 2026). [PYMNTS, "Anthropic On Track for First Operating Profit as Revenue Surges"](https://www.pymnts.com/?p=3752527) (May 2026). [Engadget, "DeepSeek permanently reduces the price of its flagship V4 model by 75 percent"](https://engadget.com/2180062/deepseek-permanently-reduces-the-price-of-its-flagship-v4-model-by-75-percent) (May 2026). [Arcade.dev, analysis of AI inference economics](https://arcade.dev/blog/ai-inference-economics) (December 2025). [OpenAI Help Center, "Codex rate card"](https://help.openai.com/en/articles/20001106-codex-rate-card). [GitHub Changelog, "GitHub Copilot is moving to usage-based billing"](https://github.blog/changelog/2026/04/27/github-copilot-is-moving-to-usage-based-billing/) (April 2026). [DevOps.com, on Anthropic pausing the Agent SDK billing change](https://devops.com/anthropic-hits-pause-on-claude-agent-sdk-billing-change-for-now/) (June 2026). [Datadog, "State of AI Engineering 2026"](https://www.datadoghq.com/state-of-ai-engineering/). [AI.cc press release, "Enterprise Token Costs Drop 67% Year-Over-Year"](https://www.einpresswire.com/article/911544568/aicc-report-enterprise-token-costs-drop-67-year-over-year-as-multi-model-ai-adoption-hits-record-high) (May 2026). [FinOps Foundation, "State of FinOps 2026"](https://data.finops.org/). [IDC, "The future of AI is model routing"](https://idc.com/resource-center/blog/the-future-of-ai-is-model-routing) (November 2025). [TechCrunch, "OpenRouter more than doubles valuation to $1.3B in a year"](https://techcrunch.com/2026/05/26/openrouter-more-than-doubles-valuation-to-1-3b-in-a-year/) (May 2026). [Palo Alto Networks, Form 10-K for the fiscal year ended July 31, 2026](https://www.sec.gov/Archives/edgar/data/0001327567/000132756726000023/panw-20260731.htm). Anthropic, OpenAI, Google, DeepSeek model pricing as of June 2026.